www.HIMSSConference.org
#smartHIT
FEBRUARY 11, 2019
ORLANDO, FL
Machine Learning to Predict Risk and
Enhance Efficiency
Agenda
• Introductions
• Project Goals & Objectives
• Project Overview
• Outcomes
• Success Mantra
Westchester Medical Center Health Network
WMCHealth is a 1,700-bed healthcare system headquartered in Valhalla, New York
• 10 Hospitals on 8 Campuses
• Level 1, Level 2 and Pediatric Trauma Centers
• Largest mental health system in New York State
• Academic Research Medical Center
• 12,000 Employees; 3,000 Attendings
Project Team & Workgroup
Deborah Viola – VP, Data
Management & Analytics, WMC
Michele Muldoon –VP
Population Health, BSCHS
Simer Sodhi – Director, Data
Management & Analytics, WMC
Craig Dickman – Director,
Population Health, BSCHS
Lauren Torres – Database
Developer, BSCHS
Taylor Larsen – Data
Scientist, Health Catalyst
Project Goals & Objectives
• Improve Pre-Discharge Program
• Improve Post Discharge Programs
• Timely Communication
• Follow Up Appointments
• More Engagement with patients post discharge
Project Overview
Problem
Care Managers, with limited resources, needed an automated and reliable
way to stratify a daily worklist of patients by risk.
Solution
Integrated Discharge & Population Health Planning Platform
We built a predictive model that creates a patient work list, stratified by risk &
paired with additional information, accessible through a care management
platform.
Defining “Success”
Improved accuracy and efficiency in identifying patients most at risk.
Project Life Cycle
Current State
Data Set
Preparation
Model
Implementation
Intervention
Execution Steps
Current State
Assessment
Identification
Acquisition
Staging
Cleaning
Transformation
Standardization
Training and
Evaluating
Delivery
Mechanism
Process
Improvement
Project Life Cycle
Current State
Data Set
Cleansing
Model
Implementation
Intervention
Execution Steps
New
Observations
New
Processes
Machine Learning Model – Data & Variable List
Clinical Data
Diagnosis
Procedure
Encounter
Record
Facility/Account
Orders
Lab Result
Medication
Demographics Data
Age / Gender
Socioeconomic
status
Ethnicity
Medical/BH History
Region
Payer
Provider
Inpatient Data – 54,000 Inpatient discharges between 1/1/2012 and 6/30/2017
Exclusions
Deceased
Patients
Left Against Medical
Advice
Cancer
Patients
Rehab Patients
Acute Care Transfers
Machine Learning – Model Description
The predictive model was developed
using:
• The random forest algorithm
leveraging 200 random trees
• And the Gini impurity index to
select the final variables for
constructing trees
Machine Learning platform – healthcare.ai
Married?
Male?
Age >
65?
LOS > 5?
Readmit
No
Readmit
No
Readmit
No
Readmit
Male?
Age >
65?
Readmit
No
Readmit
Age >
65?
LOS > 5?
Readmit
No
Readmit
No
Readmit
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
No
No
NoNo
No
No
Example of one Decision Tree
Machine Learning Model – Data & Variable List
Demographics
Age at admit
Financial class
Marital status
Visit details
Admit Source
Admit Type
Discharge Disposition
Total Variable Cost
Acuity
MSDRG Category
MSDRG Weight
ED/IP Utilization
Excess Acute care Days
Excess Acute care Visits
Historical Readmits
6 Months ED/IP visits
12 months ED/IP visits
Length of stay (LOS)
Average historical LOS
Current LOS
Total historical LOS
Comorbidities
Charlson-Deyo score
Elixhauser score
History of COPD
History of Depression
History of Diabetes
Machine Learning Model - Accuracy
The predictive model we developed is more
accurate, resulting in fewer false positives and
more true positives.
Locally Trained Model
Extensive Variable List
Complete Data Set
72%
62%
61%
28%
38%
39%
CUSTOM
MODEL
LACE EMR
% of Predictions Correctly
Classified
as High and Low Risk
17%
increase in
accuracy
Implementation
Fully Automated Solution
Exhaustive list of
Discharge Data elements
Extensive Selection Criteria
Embedded within Care
Management platform
Implementation
Training
Action
Evaluation
Learning
Training Care Managers and Care Coordinators
Care Planning Processes
Periodic Review Meeting
Better understanding of Population and Case Mix
Persistent Maintenance
Current State
Data Set
Cleansing
Model
Implementation
Intervention
Execution Steps
Training
Action
Evaluation
Learning
Outcomes
Savings of 1,327 hours
per year
Improved Organizational
Efficiency
Standardization of
Workflows - Faster
Follow-ups
Data-driven decisions to
connect patients with support
services
Success MANTRA
• Multifaceted team
• Model trained on local data
• Workflow integration
• Simplicity & Transparency
• Persistent Maintenance
www.HealthcareMachineLearningAI.com
#smartHIT
Simer Sodhi, MS, MBA
Director , Data Management & Analytics
Westchester Medical Center, New York
@simarsodhi17
